What are AI agents, and should you use one?
A chatbot answers your question. An agent goes off and does the task, clicking, typing and filling in forms. Here’s what they’re good for, and how to use one safely.
Video transcript
You’ve probably heard that AI agents are the next big thing. So what are they, what can they actually do, and should you use one?
Here’s the difference. A chatbot answers your question. An agent acts: it searches the web, clicks through sites, fills in forms, and can even write and run code.
You give it a goal. It makes a plan, takes a step, checks what happened, and goes round again until the job is done.
Start with jobs where a mistake costs you nothing, like researching a topic into a report, comparing products before you buy, or tidying a messy folder of files.
But know the risks. Small mistakes snowball over many steps. It needs access to your accounts. Web pages can hide instructions to trick it. And long tasks can be slow and expensive.
So start small. Pick low-stakes tasks and watch it work. Make it ask before it sends, buys or deletes anything, and give it only the access it needs.
And if it’s a task you repeat the same way each time, a simple automation in Zapier or Make is often faster, cheaper and more reliable than an agent.
Read the full guide below for good first jobs, and a brief you can give your first agent.
In 30 seconds
- A chatbot answers; an agent acts, working through a task step by step and checking as it goes.
- Start with low-stakes jobs: research into a report, comparing products or tidying files.
- Watch it work, give it the least access it needs, and approve anything that sends, buys or deletes.
Ask a chatbot how to find a cheap train to Edinburgh and you get good advice. Ask an Jargon busterAI agent: An AI that carries out a task in several steps, such as searching, clicking and filling in forms, rather than just replying. and it opens the booking sites, compares times and fares, and comes back with three options for you to choose from.
Agents are genuinely useful, but they’re slower and less reliable than the demos suggest. The trick is knowing which jobs to hand over, and how to stay in control.
Chatbot or agent?
| Chatbot | Agent |
|---|---|
| Answers in one reply | Works through a task in many steps |
| Tells you what to do | Does it: searches, clicks, types and fills in forms |
| Hands you text to copy | Can write and run code, and create or change files |
| A mistake stays on the screen | A mistake can become an action |
- GoalYou describe the outcome you want
- PlanIt breaks the job into steps
- ActIt searches, clicks, types or runs code
- CheckIt looks at the result and decides what next
Agent features are turning up in assistants, web browsers and workplace apps. Some take over a browser while you watch; others work in the background and report back when they’re done.
Good first jobs for an agent
- Research into a report“Read the best sources on heat pumps for a 1930s semi and write me a two-page summary, with links.”
- Compare products“Compare five family tents on size, weight, reviews and price, and put them in a table.”
- Tidy your files“Rename these holiday photos by date and place, and sort them into a folder for each trip.”
What they have in common: if the agent gets it wrong, you’ve lost a little time, not money or data. Ask for its sources, and for files, let it work on a copy.
Goal: [what you want at the end, such as a one-page comparison of five family tents]. Use [at least five] reputable sources and link each one. Don’t sign in, buy anything, fill in forms or download files. If a site asks for personal details, stop and ask me. Put the results in a table with [the columns you want], then recommend one and say why. Finish with a list of anything you couldn’t check.
The risks, honestly
- Small mistakes compound. One wrong turn early on can send every later step off course, and the final report can still sound confident.
- It needs access. To act for you, an agent may need your browser, email, files or logins. Anything it can reach, it can get wrong.
- Hidden instructions. Web pages and emails can contain text written to hijack an agent, a trick called Jargon busterPrompt injection: Hidden instructions in a web page or document that try to trick an AI into doing something you didn’t ask.. It’s why you approve anything important yourself.
- Slow and sometimes costly. Long tasks take many minutes and far more computing than a chat, so you may hit usage limits or pay extra.
How to start safely
- Start low-stakes. Stick to jobs like the three above while you learn how it behaves.
- Watch it work. Agents usually show each step. Stay with the first few runs and stop it if it wanders off.
- Approve the big moves. Make it ask before it sends, buys, books or deletes anything, and read what it’s about to do before you say yes.
- Give it the least access it needs. Use a separate browser profile or account with nothing sensitive in it, and connect only the apps the job requires.
- Check the result. Open the links, spot-check the numbers and look over any files it changed.
New to all this? Start with the AI tools worth trying, and read our guide to staying safe with AI before you connect an agent to your accounts.
Check yourself
3 quick questions nothing is savedTools in this guide
Spotted a mistake? Tell us and an editor will check it.
